muonroi-cli 1.7.2 → 1.8.0

package/dist/src/pil/__tests__/layer3-ee-injection.test.js

@@ -1,4 +1,5 @@
 import { beforeEach, describe, expect, test, vi } from "vitest";
+import { sessionRecallLedger } from "../../ee/recall-ledger.js";
 import { layer3EeInjection, RECALL_FEEDBACK_NUDGE } from "../layer3-ee-injection.js";
 vi.mock("../../ee/bridge.js", () => ({
     searchByText: vi.fn().mockResolvedValue([]),
@@ -22,6 +23,9 @@ describe("layer3EeInjection (bridge-based)", () => {
     beforeEach(() => {
         vi.mocked(searchByText).mockReset();
         vi.mocked(searchByText).mockResolvedValue([]);
+        // The recall-feedback ledger is a process singleton; reset so passive-injection
+        // debt recorded by one test does not lengthen the pending reminder in the next.
+        sessionRecallLedger.reset();
     });
     test("Test 1: returns enriched context with hints when searchByText returns points with payload.text", async () => {
         vi.mocked(searchByText).mockResolvedValue([
@@ -93,10 +97,10 @@ describe("layer3EeInjection (bridge-based)", () => {
                 const chars = parseInt(charsMatch[1], 10);
                 // Two parallel collections, each at 15% of budget: 15% of 100 tokens * 4 chars/token
                 // = 60 chars per block + 3 for "..." suffix, joined with newline. Allow generous
-                // ceiling for header text + 2 blocks, PLUS the fixed ee_feedback nudge appended
-                // when rateable experience is present. The 2000-char input means a truncation
-                // regression would blow well past this bound regardless.
-                expect(chars).toBeLessThanOrEqual(260 + RECALL_FEEDBACK_NUDGE.length + 1);
+                // ceiling for header text + 2 blocks, PLUS the dynamic pending-feedback reminder
+                // (≤5 [id collection] lines) that replaced the fixed nudge. The 2000-char input
+                // means a truncation regression would blow well past this bound regardless.
+                expect(chars).toBeLessThanOrEqual(600);
             }
         }
     });
@@ -137,6 +141,65 @@ describe("Layer 3 formatter mode (ctx._brainData populated)", () => {
         expect(result.enriched).toContain("mock fs in unit tests");
         expect(result.t1Rules).toEqual(["never skip tests"]);
     });
+    test("unified path: records rateable points (with id) into the ledger + names [id collection] in the reminder", async () => {
+        // Symmetric with the legacy-path recall-loop closure. When the server (PIL
+        // schema_version 1.1+) attributes points with id/collection, the unified
+        // formatter must record them as pending debt and surface an actionable reminder.
+        sessionRecallLedger.reset();
+        const { layer3EeInjection } = await import("../layer3-ee-injection.js");
+        const ctx = {
+            raw: "x",
+            enriched: "x",
+            taskType: "debug",
+            domain: null,
+            confidence: 0.85,
+            outputStyle: "balanced",
+            tokenBudget: 2000,
+            metrics: null,
+            layers: [],
+            _brainData: {
+                t0_principles: [{ text: "always run tests", score: 0.9, id: "pr1", collection: "experience-principles" }],
+                t1_rules: [],
+                t2_patterns: [{ text: "mock fs in unit tests", score: 0.7, id: "be1", collection: "experience-behavioral" }],
+                retrieval_skipped_reason: null,
+            },
+        };
+        const result = await layer3EeInjection(ctx);
+        // Inline [id:..] handle rendered so the reminder refers to something visible.
+        expect(result.enriched).toContain("[id:pr1]");
+        // Dynamic reminder names the actual [id collection] so ee_feedback is actionable.
+        expect(result.enriched).toMatch(/ee_feedback\(id, collection, followed\|ignored\|noise\)/);
+        expect(result.enriched).toContain("[pr1 experience-principles]");
+        expect(result.enriched).toContain("[be1 experience-behavioral]");
+        // Both points recorded as rateable pending debt.
+        expect(sessionRecallLedger.pendingCount()).toBe(2);
+    });
+    test("unified path: no id (older server) renders text but stays unrateable (static nudge)", async () => {
+        sessionRecallLedger.reset();
+        const { layer3EeInjection } = await import("../layer3-ee-injection.js");
+        const ctx = {
+            raw: "x",
+            enriched: "x",
+            taskType: "debug",
+            domain: null,
+            confidence: 0.85,
+            outputStyle: "balanced",
+            tokenBudget: 2000,
+            metrics: null,
+            layers: [],
+            _brainData: {
+                t0_principles: [{ text: "always run tests", score: 0.9 }],
+                t1_rules: [],
+                t2_patterns: [{ text: "mock fs in unit tests", score: 0.7 }],
+                retrieval_skipped_reason: null,
+            },
+        };
+        const result = await layer3EeInjection(ctx);
+        expect(result.enriched).toContain("always run tests");
+        // No ids → nothing recorded; the static nudge is the fallback.
+        expect(sessionRecallLedger.pendingCount()).toBe(0);
+        expect(result.enriched).toContain(RECALL_FEEDBACK_NUDGE);
+    });
     test("emits no block when ctx._brainData is null AND legacy disabled by flag", async () => {
         const { layer3EeInjection } = await import("../layer3-ee-injection.js");
         const ctx = {

package/dist/src/pil/__tests__/layer3-injected-chunk.test.js

@@ -24,6 +24,7 @@ vi.mock("../../storage/interaction-log.js", () => ({
 vi.mock("../../ee/intercept.js", () => ({
     updateLastSurfacedState: vi.fn(),
 }));
+import { sessionRecallLedger } from "../../ee/recall-ledger.js";
 import { layer3EeInjection } from "../layer3-ee-injection.js";
 const BASE_CTX = {
     raw: "test prompt for experience injection",
@@ -41,6 +42,9 @@ describe("layer3 experience_injected chunk emission (CQ-16b)", () => {
     beforeEach(() => {
         capturedSinkCalls.length = 0;
         mockSearchByText.mockResolvedValue([]);
+        // Reset the process-singleton recall ledger so pending debt from a prior test
+        // does not change the dynamic feedback reminder content of the next.
+        sessionRecallLedger.reset();
     });
     it("emits experience_injected chunk when searchByText returns high-score point", async () => {
         mockSearchByText.mockResolvedValue([
@@ -103,8 +107,16 @@ describe("layer3 experience_injected chunk emission (CQ-16b)", () => {
                 collection: "experience-behavioral",
             },
         ]);
-        const result = await layer3EeInjection(BASE_CTX);
-        expect(result.enriched).toContain("ee_feedback(id, followed|ignored|noise)");
+        const result = await layer3EeInjection({ ...BASE_CTX, sessionId: "test-session-nudge" });
+        // The dynamic pending-feedback reminder replaced the fixed nudge when the ledger
+        // is enabled (default soft); it names the actual [id collection] so ee_feedback is
+        // actionable. Accept either form (reminder when the ledger recorded debt, or the
+        // static fallback when disabled).
+        expect(result.enriched).toMatch(/ee_feedback\(id, (collection, )?followed\|ignored\|noise\)/);
+        // Reminder names the actual [id collection]; the mock returns the same point in
+        // both search arms so the principles arm wins first-sighting (real searches return
+        // distinct points per collection) — assert id-named, collection-agnostic.
+        expect(result.enriched).toMatch(/\[p1 experience-(principles|behavioral)\]/);
     });
     it("does NOT emit experience_injected when searchByText returns empty array", async () => {
         mockSearchByText.mockResolvedValue([]);

package/dist/src/pil/__tests__/layer4-gsd.test.js

@@ -1,7 +1,11 @@
 import { describe, expect, it, vi } from "vitest";
 vi.mock("../../ee/bridge.js", () => ({
     routeTask: vi.fn().mockResolvedValue(null),
+    // WhoAmI v4.0: L4 reads delegation_style to bias a 'discuss' phase. Default null
+    // = fail-open, so every pre-existing test below behaves exactly as before.
+    getWhoAmIProfile: vi.fn(() => null),
 }));
+import { getWhoAmIProfile } from "../../ee/bridge.js";
 import { layer4Gsd } from "../layer4-gsd.js";
 function makeCtx(overrides = {}) {
     return {
@@ -139,6 +143,32 @@ describe("layer4Gsd (playbook)", () => {
         const result = await layer4Gsd(makeCtx({ raw, enriched: raw, taskType: "refactor", intentKind: "task", deliverableKind: "code" }));
         expect(result.enriched).not.toContain("QUESTION / explanatory");
     });
+    it("autonomous delegation drops a 'discuss' phase (WhoAmI v4.0 bias toward direct)", async () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValueOnce({
+            level: "standard",
+            dims: { "work_patterns.delegation_style": { value: "autonomous", confidence: 0.68, samples: 19 } },
+        });
+        const result = await layer4Gsd(makeCtx({ gsdPhase: "discuss", raw: "do the thing" }));
+        expect(result.gsdPhase ?? null).toBeNull();
+        const layer = result.layers.find((l) => l.name === "gsd-workflow-structuring");
+        expect(layer.delta).toContain("phase=none");
+        expect(layer.delta).toContain("autonomous:skip-discuss");
+    });
+    it("collaborative delegation keeps the 'discuss' phase", async () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValueOnce({
+            level: "standard",
+            dims: { "work_patterns.delegation_style": { value: "collaborative", confidence: 0.7, samples: 20 } },
+        });
+        const result = await layer4Gsd(makeCtx({ gsdPhase: "discuss", raw: "do the thing" }));
+        expect(result.gsdPhase).toBe("discuss");
+        const layer = result.layers.find((l) => l.name === "gsd-workflow-structuring");
+        expect(layer.delta).toContain("phase=discuss");
+    });
+    it("absent profile leaves 'discuss' untouched (fail-open default)", async () => {
+        // getWhoAmIProfile defaults to null in the mock — the phase must be unchanged.
+        const result = await layer4Gsd(makeCtx({ gsdPhase: "discuss", raw: "do the thing" }));
+        expect(result.gsdPhase).toBe("discuss");
+    });
     it("uses ctx.gsdPhase from L1 (unified path) without calling routeTask", async () => {
         const { routeTask } = await import("../../ee/bridge.js");
         vi.mocked(routeTask).mockClear();

package/dist/src/pil/__tests__/layer5-context.test.js

@@ -3,6 +3,9 @@ vi.mock("../../ee/bridge.js", () => ({
     getEmbeddingRaw: vi.fn().mockResolvedValue(null),
     searchCollection: vi.fn().mockResolvedValue([]),
     searchByText: vi.fn().mockResolvedValue([]),
+    // WhoAmI v4.0: L5 reads session_length to tune the resume-digest stale window.
+    // Default null = fail-open → 30m default, so the pre-v4.0 assertions below hold.
+    getWhoAmIProfile: vi.fn(() => null),
 }));
 vi.mock("node:fs", async (importOriginal) => {
     const actual = await importOriginal();
@@ -16,7 +19,8 @@ vi.mock("node:fs", async (importOriginal) => {
         },
     };
 });
-import { layer5Context } from "../layer5-context.js";
+import { getWhoAmIProfile } from "../../ee/bridge.js";
+import { layer5Context, staleThresholdMsForSessionLength } from "../layer5-context.js";
 function makeCtx(overrides = {}) {
     return {
         raw: "continue working on the auth module",
@@ -133,5 +137,39 @@ describe("layer5Context", () => {
         expect(result.enriched).toContain("already enriched prompt");
         expect(result.enriched).toContain("[flow-context:");
     });
+    // WhoAmI v4.0: session_length tunes the resume-digest stale window. A 45-minute-old
+    // digest is stale at the 30m default but FRESH for a long-session user (window→60m).
+    // This proves the profile actually flows through layer5Context, not just the helper.
+    it("long session_length relaxes the stale window — 45m digest is not stale", async () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValueOnce({
+            level: "minimal",
+            dims: { "work_patterns.session_length": { value: "long", confidence: 0.5, samples: 16 } },
+        });
+        const result = await layer5Context(makeCtx({ resumeDigest: "Deep-work session context", digestAgeMs: 45 * 60 * 1000 }));
+        expect(result.enriched).not.toContain("stale");
+        const layer = result.layers.find((l) => l.name === "context-enrichment");
+        expect(layer.delta).toContain("window=60m");
+        expect(layer.delta).not.toContain(",stale");
+    });
+    it("short session_length tightens the stale window — 20m digest is stale", async () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValueOnce({
+            level: "minimal",
+            dims: { "work_patterns.session_length": { value: "short", confidence: 0.5, samples: 16 } },
+        });
+        const result = await layer5Context(makeCtx({ resumeDigest: "Quick-burst session context", digestAgeMs: 20 * 60 * 1000 }));
+        expect(result.enriched).toContain("stale");
+        const layer = result.layers.find((l) => l.name === "context-enrichment");
+        expect(layer.delta).toContain("window=15m");
+    });
+});
+describe("staleThresholdMsForSessionLength — pure session_length → stale window mapping", () => {
+    it("long → 60m, short → 15m, medium/absent/unknown → 30m default (fail-open)", () => {
+        expect(staleThresholdMsForSessionLength("long")).toBe(60 * 60 * 1000);
+        expect(staleThresholdMsForSessionLength("short")).toBe(15 * 60 * 1000);
+        expect(staleThresholdMsForSessionLength("medium")).toBe(30 * 60 * 1000);
+        expect(staleThresholdMsForSessionLength(undefined)).toBe(30 * 60 * 1000);
+        expect(staleThresholdMsForSessionLength(null)).toBe(30 * 60 * 1000);
+        expect(staleThresholdMsForSessionLength("telegraphic")).toBe(30 * 60 * 1000);
+    });
 });
 //# sourceMappingURL=layer5-context.test.js.map

package/dist/src/pil/__tests__/layer6-output.test.js

@@ -1,8 +1,11 @@
-import { beforeEach, describe, expect, it, vi } from "vitest";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { getWhoAmIProfile } from "../../ee/bridge.js";
 import { applyPilSuffix, getResponseToolSet, isImplementationIntent, isQuestionLike, layer6Output, } from "../layer6-output.js";
-// Mock bridge for PIL-03 classifyViaBrain tests
+// Mock bridge for PIL-03 classifyViaBrain tests. getWhoAmIProfile defaults to null
+// (fail-open) so applyPilSuffix never reads the real on-device profile during tests.
 vi.mock("../../ee/bridge.js", () => ({
     classifyViaBrain: vi.fn().mockResolvedValue(null),
+    getWhoAmIProfile: vi.fn(() => null),
 }));
 async function getMockBrain() {
     const { classifyViaBrain } = await import("../../ee/bridge.js");
@@ -434,4 +437,29 @@ describe("isImplementationIntent — improve / cải thiện (regression: sessio
         expect(isImplementationIntent("why does the suite fail — break it down")).toBe(false);
     });
 });
+describe("applyPilSuffix — WhoAmI feedback_style correction-affordance", () => {
+    const withFeedback = (value) => ({
+        level: "standard",
+        dims: { "communication.feedback_style": { value, confidence: 0.7, samples: 20 } },
+    });
+    // Reset around each test so a mockReturnValue never leaks into (or out of) this
+    // block — clearAllMocks elsewhere only clears history, not the return override.
+    beforeEach(() => vi.mocked(getWhoAmIProfile).mockReturnValue(null));
+    afterEach(() => vi.mocked(getWhoAmIProfile).mockReturnValue(null));
+    it("precise-correction + action task (refactor) → AUDITABLE CHANGES rule", () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValue(withFeedback("precise-correction"));
+        expect(applyPilSuffix("SYS", makeCtx("refactor"))).toContain("AUDITABLE CHANGES");
+    });
+    it("precise-correction + non-action task (analyze) → no rule (diffs don't apply)", () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValue(withFeedback("precise-correction"));
+        expect(applyPilSuffix("SYS", makeCtx("analyze"))).not.toContain("AUDITABLE CHANGES");
+    });
+    it("implicit feedback → no rule (intentionally unwired — covered by brevity/delegation)", () => {
+        vi.mocked(getWhoAmIProfile).mockReturnValue(withFeedback("implicit"));
+        expect(applyPilSuffix("SYS", makeCtx("debug"))).not.toContain("AUDITABLE CHANGES");
+    });
+    it("absent profile → no rule (fail-open)", () => {
+        expect(applyPilSuffix("SYS", makeCtx("generate"))).not.toContain("AUDITABLE CHANGES");
+    });
+});
 //# sourceMappingURL=layer6-output.test.js.map

package/dist/src/pil/__tests__/orchestrator-integration.test.js

@@ -17,6 +17,8 @@ vi.mock("../../ee/bridge.js", () => ({
     searchCollection: vi.fn().mockResolvedValue([]),
     getEmbeddingRaw: vi.fn().mockResolvedValue(null),
     routeTask: vi.fn().mockResolvedValue(null),
+    getWhoAmIProfile: vi.fn(() => null),
+    outputStyleFromProfile: vi.fn(() => null),
 }));
 import { applyPilSuffix, runPipeline } from "../index.js";
 beforeEach(() => {
@@ -27,7 +29,7 @@ describe("PIL orchestrator contract", () => {
         const ctx = await runPipeline("refactor this function");
         // Layer 2 personality now appends hints, so enriched starts with raw
         expect(ctx.enriched.startsWith("refactor this function")).toBe(true);
-        expect(ctx.layers).toHaveLength(6);
+        expect(ctx.layers).toHaveLength(7);
     });
     it("applyPilSuffix appends OUTPUT RULES suffix for coding taskType (non-null)", () => {
         const codingCtx = {

package/dist/src/pil/__tests__/pipeline.test.js

@@ -8,6 +8,8 @@ vi.mock("../../ee/bridge.js", () => ({
     searchCollection: vi.fn().mockResolvedValue([]),
     getEmbeddingRaw: vi.fn().mockResolvedValue(null),
     routeTask: vi.fn().mockResolvedValue(null),
+    getWhoAmIProfile: vi.fn(() => null),
+    outputStyleFromProfile: vi.fn(() => null),
 }));
 import { classify } from "../../router/classifier/index.js";
 import { runPipeline } from "../pipeline.js";
@@ -18,10 +20,10 @@ beforeEach(() => {
     mockClassify.mockReturnValue({ tier: "hot", confidence: 0.85, reason: "regex:refactor" });
 });
 describe("runPipeline()", () => {
-    it("returns PipelineContext with 6 LayerResults for normal input", async () => {
+    it("returns PipelineContext with 7 LayerResults for normal input", async () => {
         const ctx = await runPipeline("refactor this function");
         expect(ctx.raw).toBe("refactor this function");
-        expect(ctx.layers).toHaveLength(6);
+        expect(ctx.layers).toHaveLength(7);
     });
     it("returns enriched that starts with raw (layers may append hints)", async () => {
         const ctx = await runPipeline("some prompt");
@@ -44,24 +46,25 @@ describe("runPipeline()", () => {
         const ctx = await runPipeline("");
         expect(ctx.raw).toBe("");
         expect(ctx.enriched.startsWith(ctx.raw)).toBe(true);
-        expect(ctx.layers).toHaveLength(6);
+        expect(ctx.layers).toHaveLength(7);
     });
     it("conversational turn (taskType=null) skips layers 2-5 with delta=skipped:null-taskType", async () => {
         mockClassify.mockReturnValue({ tier: "abstain", confidence: 0.2, reason: "low-confidence" });
         const ctx = await runPipeline("hello how are you");
-        expect(ctx.layers).toHaveLength(6);
+        expect(ctx.layers).toHaveLength(7);
         expect(ctx.layers[1].delta).toBe("skipped:null-taskType");
         expect(ctx.layers[2].delta).toBe("skipped:null-taskType");
         expect(ctx.layers[3].delta).toBe("skipped:null-taskType");
         expect(ctx.layers[4].delta).toBe("skipped:null-taskType");
+        expect(ctx.layers[5].delta).toBe("skipped:null-taskType");
         expect(ctx.taskType).toBeNull();
     });
-    it("coding task runs all 6 layers normally (no skip)", async () => {
+    it("coding task runs all 7 layers normally (no skip)", async () => {
         const ctx = await runPipeline("refactor this function");
-        expect(ctx.layers).toHaveLength(6);
+        expect(ctx.layers).toHaveLength(7);
         expect(ctx.taskType).toBe("refactor");
-        // layers 2-5 should NOT have skipped delta
-        for (let i = 1; i <= 4; i++) {
+        // layers 2-5 (which include indices 1 to 5) should NOT have skipped delta
+        for (let i = 1; i <= 5; i++) {
             expect(ctx.layers[i].delta).not.toBe("skipped:null-taskType");
         }
     });

package/dist/src/pil/__tests__/response-tools.test.js

@@ -172,6 +172,7 @@ describe("schemas validate correct input", () => {
         expect(result.success).toBe(false);
     });
 });
+import { normalizeStructuredResponseTaskType } from "../response-tools.js";
 describe("respond_general catch-all tool", () => {
     it("respond_general tool exists with correct schema", () => {
         const tools = buildResponseTools("general");
@@ -187,4 +188,55 @@ describe("respond_general catch-all tool", () => {
         expect(isResponseTool("respond_general")).toBe(true);
     });
 });
+describe("normalizeStructuredResponseTaskType", () => {
+    // Regression: session 48d22fe436f6 — model called respond_analyze but sent
+    // { response: "..." } (general shape). TUI rendered empty findings box.
+    it("normalizes 'analyze' → 'general' when findings missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("analyze", { response: "Some analysis text" })).toBe("general");
+    });
+    it("keeps 'analyze' when findings array is present", () => {
+        expect(normalizeStructuredResponseTaskType("analyze", {
+            findings: [{ text: "issue", evidence: "file:1", severity: "high" }],
+        })).toBe("analyze");
+    });
+    it("normalizes 'debug' → 'general' when root_cause missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("debug", { response: "Bug is in the store" })).toBe("general");
+    });
+    it("keeps 'debug' when root_cause is present", () => {
+        expect(normalizeStructuredResponseTaskType("debug", {
+            hypothesis: "race condition",
+            root_cause: "shared mutable state",
+            fix: { file: "src/a.ts", diff: "+x" },
+            verify_command: "bun test",
+        })).toBe("debug");
+    });
+    it("normalizes 'plan' → 'general' when steps missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("plan", { response: "Step 1: do A" })).toBe("general");
+    });
+    it("normalizes 'refactor' → 'general' when changes missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("refactor", { response: "Refactored X" })).toBe("general");
+    });
+    it("normalizes 'documentation' → 'general' when content missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("documentation", { response: "Docs here" })).toBe("general");
+    });
+    it("normalizes 'generate' → 'general' when files missing but response present", () => {
+        expect(normalizeStructuredResponseTaskType("generate", { response: "Generated code" })).toBe("general");
+    });
+    it("keeps 'general' unchanged always", () => {
+        expect(normalizeStructuredResponseTaskType("general", { response: "hello" })).toBe("general");
+        expect(normalizeStructuredResponseTaskType("general", {})).toBe("general");
+    });
+    it("does NOT normalize when response field is empty string", () => {
+        expect(normalizeStructuredResponseTaskType("analyze", { response: "" })).toBe("analyze");
+    });
+    it("does NOT normalize when response field is whitespace-only", () => {
+        expect(normalizeStructuredResponseTaskType("analyze", { response: "   " })).toBe("analyze");
+    });
+    it("does NOT normalize when response field is missing entirely", () => {
+        expect(normalizeStructuredResponseTaskType("analyze", { summary: "Some summary" })).toBe("analyze");
+    });
+    it("keeps unknown taskType unchanged", () => {
+        expect(normalizeStructuredResponseTaskType("unknown_type", { response: "hello" })).toBe("unknown_type");
+    });
+});
 //# sourceMappingURL=response-tools.test.js.map

package/dist/src/pil/agent-operating-contract.d.ts

@@ -37,7 +37,7 @@
  * one imperative line targeting that phase's most damaging failure mode. Kept
  * tight (primacy matters more than detail; tokens are the cost).
  */
-export declare const AGENT_OPERATING_CONTRACT = "[AGENT OPERATING CONTRACT \u2014 read first; applies to every step]\n\n1. BEFORE ACTING: do only what was asked. Never assume scope or facts \u2014 if ambiguous, ask or use defaults; never invent requirements.\n2. READING: base statements on what you read/ran THIS turn. Do not infer contents of files you did not open.\n3. EXECUTING: smallest correct change; never widen scope or mask failures (no `|| true`, skipped tests, or swallowed catch).\n4. WHEN UNSURE: verify and cross-check BEFORE concluding. Bugs need a reproduction; reading code is not proof.\n5. REPORTING: answer ONLY what was asked. Every fact or file:line MUST come from this turn; else label \"unverified\"; do not guess. Synthesize evidence gracefully \u2014 do NOT dump massive verbatim tool outputs into the final answer. Cite concise file:line references. Never claim a build/test ran, or describe edits, you did not actually do this turn; if a check can't run, fix it or say so \u2014 don't imply success.\n\n6. LANGUAGE: Reply in user's detected language for final output. Internal reasoning, tools, and code remain in English.\n\n7. ANTI-M\u00D9 / COMPACTION: After seeing \"[pre-compaction warning at step...\" or \"[context compacted at step...\", decide if you need full prior tool results. Emit PRESERVE_FULL_CONTEXT for full veto this turn, or the lighter KEEP_TOOL_IDS: id1,id2 (ids from prior stub \"(id=...)\") to protect only high-value results (read_file/grep on src/PLAN/error etc are auto-protected). Use the ee_query tool with \"tool-artifact id=XXX\" for on-demand full re-hydrate of elided ones. Self-check \"task finished?\" / \"compacted yet?\". Use EE checkpoints.\n\n8. GIT SAFETY: never push on red \u2014 run the check, await its result in a SEPARATE step, confirm 0 failures, then push. Never `git add -A`/`commit -a`; stage explicitly so secrets (.env, .muonroi-cli/, keys) aren't committed. Never `--no-verify`.\n\n[END CONTRACT \u2014 instructions follow]";
+export declare const AGENT_OPERATING_CONTRACT = "[AGENT OPERATING CONTRACT \u2014 read first; applies to every step]\n\n1. BEFORE ACTING: do only what was asked. Never assume scope or facts \u2014 if ambiguous, ask or use defaults; never invent requirements. RESEARCH FIRST: explore code (read/grep) and recall EE brain before editing. RECALL FIRST: ee.query in unfamiliar areas to surface past lessons.\n2. READING: base statements on what you read/ran THIS turn. Do not infer contents of files you did not open.\n3. EXECUTING: smallest correct change; never widen scope or mask failures (no `|| true`, skipped tests, or swallowed catch).\n4. WHEN UNSURE: verify and cross-check BEFORE concluding. Reading code is not proof \u2014 reproduce the bug.\n5. REPORTING: answer ONLY what was asked. Every fact or file:line MUST come from this turn; else label \"unverified\"; do not guess. Synthesize evidence gracefully \u2014 do NOT dump massive verbatim tool outputs into the final answer. Cite concise file:line references. Never claim a build/test ran, or describe edits, you did not actually do this turn; if a check can't run, fix it or say so \u2014 don't imply success.\n\n6. LANGUAGE: Reply in user's detected language for final output. Internal reasoning, tools, and code remain in English.\n\n7. ANTI-M\u00D9 / COMPACTION: On compaction, emit PRESERVE_FULL_CONTEXT (veto) or KEEP_TOOL_IDS (from stub id=) to protect results. Use ee_query with \"tool-artifact id=XXX\" to re-hydrate. Self-check via EE checkpoints. Suggest /compact near tool limits.\n\n8. GIT SAFETY: never push on red \u2014 run the check, await its result in a SEPARATE step, confirm 0 failures, then push. Never `git add -A`/`commit -a`; stage explicitly so secrets (.env, .muonroi-cli/, keys) aren't committed. Never `--no-verify`.\n\n9. VERIFICATION: when finishing a task, ALWAYS self-verify your work. Use the `selfverify_*` native tools (start/status/result) to run the QA harness which drives the live TUI like a real user to catch regressions that unit tests can't.\n\n[END CONTRACT \u2014 instructions follow]";
 export interface ContractSectionOptions {
     /** Chitchat turns carry no tools and make no factual claims — skip the contract. */
     chitchat?: boolean;

package/dist/src/pil/agent-operating-contract.js

@@ -39,18 +39,20 @@
  */
 export const AGENT_OPERATING_CONTRACT = `[AGENT OPERATING CONTRACT — read first; applies to every step]
 
-1. BEFORE ACTING: do only what was asked. Never assume scope or facts — if ambiguous, ask or use defaults; never invent requirements.
+1. BEFORE ACTING: do only what was asked. Never assume scope or facts — if ambiguous, ask or use defaults; never invent requirements. RESEARCH FIRST: explore code (read/grep) and recall EE brain before editing. RECALL FIRST: ee.query in unfamiliar areas to surface past lessons.
 2. READING: base statements on what you read/ran THIS turn. Do not infer contents of files you did not open.
 3. EXECUTING: smallest correct change; never widen scope or mask failures (no \`|| true\`, skipped tests, or swallowed catch).
-4. WHEN UNSURE: verify and cross-check BEFORE concluding. Bugs need a reproduction; reading code is not proof.
+4. WHEN UNSURE: verify and cross-check BEFORE concluding. Reading code is not proof — reproduce the bug.
 5. REPORTING: answer ONLY what was asked. Every fact or file:line MUST come from this turn; else label "unverified"; do not guess. Synthesize evidence gracefully — do NOT dump massive verbatim tool outputs into the final answer. Cite concise file:line references. Never claim a build/test ran, or describe edits, you did not actually do this turn; if a check can't run, fix it or say so — don't imply success.
 
 6. LANGUAGE: Reply in user's detected language for final output. Internal reasoning, tools, and code remain in English.
 
-7. ANTI-MÙ / COMPACTION: After seeing "[pre-compaction warning at step..." or "[context compacted at step...", decide if you need full prior tool results. Emit PRESERVE_FULL_CONTEXT for full veto this turn, or the lighter KEEP_TOOL_IDS: id1,id2 (ids from prior stub "(id=...)") to protect only high-value results (read_file/grep on src/PLAN/error etc are auto-protected). Use the ee_query tool with "tool-artifact id=XXX" for on-demand full re-hydrate of elided ones. Self-check "task finished?" / "compacted yet?". Use EE checkpoints.
+7. ANTI-MÙ / COMPACTION: On compaction, emit PRESERVE_FULL_CONTEXT (veto) or KEEP_TOOL_IDS (from stub id=) to protect results. Use ee_query with "tool-artifact id=XXX" to re-hydrate. Self-check via EE checkpoints. Suggest /compact near tool limits.
 
 8. GIT SAFETY: never push on red — run the check, await its result in a SEPARATE step, confirm 0 failures, then push. Never \`git add -A\`/\`commit -a\`; stage explicitly so secrets (.env, .muonroi-cli/, keys) aren't committed. Never \`--no-verify\`.
 
+9. VERIFICATION: when finishing a task, ALWAYS self-verify your work. Use the \`selfverify_*\` native tools (start/status/result) to run the QA harness which drives the live TUI like a real user to catch regressions that unit tests can't.
+
 [END CONTRACT — instructions follow]`;
 /**
  * Build the contract block for insertion at the front of the system prompt.

package/dist/src/pil/agent-operating-contract.test.js

@@ -47,8 +47,8 @@ describe("AGENT_OPERATING_CONTRACT", () => {
         expect(AGENT_OPERATING_CONTRACT).toMatch(/push on red|never push/i);
         expect(AGENT_OPERATING_CONTRACT).toMatch(/git add -A|stage explicitly/i);
     });
-    it("stays compact (under 1900 chars) to preserve attention budget on every turn (git-safety rule added)", () => {
-        expect(AGENT_OPERATING_CONTRACT.length).toBeLessThan(1900);
+    it("stays compact (under 2000 chars) to preserve attention budget on every turn (git-safety rule added)", () => {
+        expect(AGENT_OPERATING_CONTRACT.length).toBeLessThan(2000);
     });
 });
 describe("buildContractSection", () => {

package/dist/src/pil/cheap-model-playbook.d.ts

@@ -27,7 +27,7 @@ import type { ShellKind } from "../utils/shell.js";
  * Wrapped with the `[CRITICAL TOOL-USE RULES ...]` marker so the model knows
  * to treat these as overrides to anything that follows.
  */
-export declare const CHEAP_MODEL_PLAYBOOK = "[CRITICAL TOOL-USE RULES \u2014 read before invoking any tool; these override defaults that follow]\n\n1. Bash output is AUTOMATICALLY cached. Every `bash` call returns a `run_id`\n   (e.g. `bash-1`) you can re-query via `bash_output_get(run_id, mode=tail|head|grep|lines)`.\n   - When you want only the last N lines: do NOT pipe `| tail -N`. Run the\n     bare command, then call `bash_output_get(run_id, mode=tail, lines=N)`.\n   - Same for `| head`, `| grep PATTERN`, `> file`. Pipes/redirects HIDE\n     the full output from the cache; `bash_output_get` reads from the cache\n     without re-running.\n   - This applies to EVERY bash call, not just retries.\n   - To VIEW a file use `read_file` (start_line/end_line) \u2014 never sed/cat a\n     file. `bash_output_get` is for COMMAND output, not files.\n\n2. Before reading more than 3 files to understand a topic, delegate to\n   `task(agent=\"explore\")`. The sub-agent returns a compressed summary;\n   you save reading tokens.\n\n3. Use the `grep` tool (ripgrep) for content search \u2014 NOT `bash` with\n   `grep` / `find` piped.\n\n4. When a tool returns `ERROR: ...`, do NOT retry the identical call.\n   Pick a different tool, change inputs meaningfully, or stop and report.\n\n5. Fix the ROOT CAUSE, never mask a failure to make it \"pass\"\n   (`continue-on-error`, swallowed try/catch, skipped/deleted test, `|| true`).\n   If a step fails from a missing secret/config, make it CONDITIONAL (skip when\n   absent) so it still runs when present \u2014 do NOT blanket-ignore it.\n\n6. For a build / CI / test failure, read the ACTUAL failure log or stack trace\n   BEFORE hypothesizing \u2014 fix the real error, not a guess from source alone.\n\n7. ANTI-M\u00D9 / COMPACTION (for long sessions): On pre-warn or \"[context compacted at step...\", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with \"tool-artifact id=XXX\" for on-demand full. Self-check \"task finished?\" / \"compacted yet?\". Use EE checkpoints.\n\n[END CRITICAL TOOL-USE RULES \u2014 your regular instructions begin below]\n\n";
+export declare const CHEAP_MODEL_PLAYBOOK = "[CRITICAL TOOL-USE RULES \u2014 read before invoking any tool; these override defaults that follow]\n\n1. Bash output is AUTOMATICALLY cached. Every `bash` call returns a `run_id`\n   (e.g. `bash-1`) you can re-query via `bash_output_get(run_id, mode=tail|head|grep|lines)`.\n   - When you want only the last N lines: do NOT pipe `| tail -N`. Run the\n     bare command, then call `bash_output_get(run_id, mode=tail, lines=N)`.\n   - Same for `| head`, `| grep PATTERN`, `> file`. Pipes/redirects HIDE\n     the full output from the cache; `bash_output_get` reads from the cache\n     without re-running.\n   - This applies to EVERY bash call, not just retries.\n   - To VIEW a file use `read_file` (start_line/end_line) \u2014 never sed/cat a\n     file. `bash_output_get` is for COMMAND output, not files.\n\n2. Before reading more than 3 files to understand a topic, delegate to\n   `task(agent=\"explore\")`. The sub-agent returns a compressed summary;\n   you save reading tokens.\n\n3. Use the `grep` tool (ripgrep) for content search \u2014 NOT `bash` with\n   `grep` / `find` piped.\n\n4. When a tool returns `ERROR: ...`, do NOT retry the identical call.\n   Pick a different tool, change inputs meaningfully, or stop and report.\n\n5. Fix the ROOT CAUSE, never mask a failure to make it \"pass\"\n   (`continue-on-error`, swallowed try/catch, skipped/deleted test, `|| true`).\n   If a step fails from a missing secret/config, make it CONDITIONAL (skip when\n   absent) so it still runs when present \u2014 do NOT blanket-ignore it.\n\n6. For a build / CI / test failure, read the ACTUAL failure log or stack trace\n   BEFORE hypothesizing \u2014 fix the real error, not a guess from source alone.\n\n7. ANTI-M\u00D9 / COMPACTION (for long sessions): On pre-warn or \"[context compacted at step...\", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with \"tool-artifact id=XXX\" for on-demand full. Self-check \"task finished?\" / \"compacted yet?\". Use EE checkpoints. If you are reaching tool/step limits in a long session, suggest the user run \"/compact\" in the chat to compress this session's history.\n\n[END CRITICAL TOOL-USE RULES \u2014 your regular instructions begin below]\n\n";
 /**
  * Predicate gating playbook injection.
  *

package/dist/src/pil/cheap-model-playbook.js

@@ -56,7 +56,7 @@ export const CHEAP_MODEL_PLAYBOOK = `[CRITICAL TOOL-USE RULES — read before in
 6. For a build / CI / test failure, read the ACTUAL failure log or stack trace
    BEFORE hypothesizing — fix the real error, not a guess from source alone.
 
-7. ANTI-MÙ / COMPACTION (for long sessions): On pre-warn or "[context compacted at step...", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with "tool-artifact id=XXX" for on-demand full. Self-check "task finished?" / "compacted yet?". Use EE checkpoints.
+7. ANTI-MÙ / COMPACTION (for long sessions): On pre-warn or "[context compacted at step...", emit PRESERVE_FULL_CONTEXT (full veto) or lighter KEEP_TOOL_IDS: id1,id2 (from stub id=) to protect specific high-value results. read_file/grep/lsp/bash on src/PLAN/error are auto-kept (idea 1). Use ee.query tool with "tool-artifact id=XXX" for on-demand full. Self-check "task finished?" / "compacted yet?". Use EE checkpoints. If you are reaching tool/step limits in a long session, suggest the user run "/compact" in the chat to compress this session's history.
 
 [END CRITICAL TOOL-USE RULES — your regular instructions begin below]
 

package/dist/src/pil/cheap-model-workbooks.d.ts

@@ -25,7 +25,7 @@ import type { TaskType } from "./types.js";
  * Universal anti-ramble convergence block — applies to every task type.
  * Kept tight; the per-task addendum below specialises it.
  */
-export declare const CHEAP_MODEL_CONVERGENCE = "[CONVERGENCE \u2014 minimise tool calls; the system prompt + tools are re-sent every call, so each extra step is expensive]\n\n- Plan the FEWEST reads you need, then read the specific file/section directly.\n  Do NOT broad-grep, re-read a file you already read, or explore \"just in case\".\n- The moment you have enough to act, STOP investigating and make the change.\n- Make the SMALLEST correct change for the request; do not widen scope.\n- Finish the action before you answer \u2014 never stop mid-step (e.g. \"I'm verifying\u2026\").\n  When done, state completion in ONE line (what changed + that it's verified);\n  no recap, no next-steps padding.\n- GROUND every claim in what you actually read or ran THIS turn: cite real\n  file:line, and never invent counts, line numbers, names, or bugs. If a number\n  (test/file count) is not verified by a command you ran, run the check or mark\n  it \"unverified\" \u2014 do NOT guess a value or assert a finding you did not observe.\n- ANTI-M\u00D9: After compaction note or pre-warn, emit PRESERVE_FULL_CONTEXT (full) or KEEP_TOOL_IDS: id1,id2 to protect high-value (auto for read_file/grep on src/PLAN/error). Use the ee_query tool with \"tool-artifact id=XXX\" for on-demand full re-hydrate. Recall checkpoints. ";
+export declare const CHEAP_MODEL_CONVERGENCE = "[CONVERGENCE \u2014 minimise tool calls; the system prompt + tools are re-sent every call, so each extra step is expensive]\n\n- Plan the FEWEST reads you need, then read the specific file/section directly.\n  Do NOT broad-grep, re-read a file you already read, or explore \"just in case\".\n- The moment you have enough to act, STOP investigating and make the change.\n- Make the SMALLEST correct change for the request; do not widen scope.\n- Finish the action before you answer \u2014 never stop mid-step (e.g. \"I'm verifying\u2026\").\n  When done, state completion in ONE line (what changed + that it's verified);\n  no recap, no next-steps padding.\n- GROUND every claim in what you actually read or ran THIS turn: cite real\n  file:line, and never invent counts, line numbers, names, or bugs. If a number\n  (test/file count) is not verified by a command you ran, run the check or mark\n  it \"unverified\" \u2014 do NOT guess a value or assert a finding you did not observe.\n- ANTI-M\u00D9: After compaction note or pre-warn, emit PRESERVE_FULL_CONTEXT (full) or KEEP_TOOL_IDS: id1,id2 to protect high-value (auto for read_file/grep on src/PLAN/error). Use the ee_query tool with \"tool-artifact id=XXX\" for on-demand full re-hydrate. Recall checkpoints. If you are reaching tool/step limits in a long session, suggest the user run \"/compact\" to compress history. ";
 /**
  * Map a sub-agent role (agentKey) to the workbook TaskType that best matches
  * its job. Sub-agents never run PIL Layer 1, so they have no classifier-derived

package/dist/src/pil/cheap-model-workbooks.js

@@ -36,7 +36,7 @@ export const CHEAP_MODEL_CONVERGENCE = `[CONVERGENCE — minimise tool calls; th
   file:line, and never invent counts, line numbers, names, or bugs. If a number
   (test/file count) is not verified by a command you ran, run the check or mark
   it "unverified" — do NOT guess a value or assert a finding you did not observe.
-- ANTI-MÙ: After compaction note or pre-warn, emit PRESERVE_FULL_CONTEXT (full) or KEEP_TOOL_IDS: id1,id2 to protect high-value (auto for read_file/grep on src/PLAN/error). Use the ee_query tool with "tool-artifact id=XXX" for on-demand full re-hydrate. Recall checkpoints. `;
+- ANTI-MÙ: After compaction note or pre-warn, emit PRESERVE_FULL_CONTEXT (full) or KEEP_TOOL_IDS: id1,id2 to protect high-value (auto for read_file/grep on src/PLAN/error). Use the ee_query tool with "tool-artifact id=XXX" for on-demand full re-hydrate. Recall checkpoints. If you are reaching tool/step limits in a long session, suggest the user run "/compact" to compress history. `;
 /**
  * Per-task-type addenda. Each is 1–2 tight lines targeting that type's most
  * common budget-model failure mode. Types not listed fall back to the

package/dist/src/pil/config.d.ts

@@ -5,6 +5,16 @@
  *   Default OFF during rollout; flip to ON after dual-run validation.
  */
 export declare function isUnifiedPilEnabled(): boolean;
+/**
+ * MUONROI_PIL_UNIFIED_BUDGET_MS: client-side timeout for the unified
+ * /api/pil-context call in Layer 1 Pass 3. The server's uncached path is
+ * classifier-bound (~1.5-3s; its own classifier timeout is 3500ms), so the
+ * legacy 1500ms budget aborted before the server could ever answer — the
+ * unified path then ALWAYS fell back to legacy, making MUONROI_PIL_UNIFIED=1 a
+ * no-op. Default 3500ms clears realistic uncached latency; cache hits return in
+ * ~1ms regardless. Clamped to [1000, 8000].
+ */
+export declare function getUnifiedPilBudgetMs(): number;
 /**
  * MUONROI_LLM_FIRST_CLASSIFY: model-first Layer-1 classification. When enabled
  * (default), the configured model classifies taskType/intentKind/style at the
@@ -14,5 +24,15 @@ export declare function isUnifiedPilEnabled(): boolean;
  * (the orchestrator does this on the main path); without it, the cascade runs.
  */
 export declare function isLlmFirstClassifyEnabled(): boolean;
+/**
+ * G3 (b1): on the model-first classify path, also fetch the unified pil-context
+ * brain so layer3 renders the rich source="unified" injection (server-attributed
+ * t0/t2 + t1_rules + the rateable ledger) instead of falling through to its
+ * legacy dense-only /api/search round-trip. Default ON; set "0" to revert to
+ * the legacy layer3 injection on the default path. Independent of
+ * MUONROI_PIL_UNIFIED (which gates the OFFLINE cascade's unified pass).
+ */
+export declare function isLlmFirstBrainEnabled(): boolean;
 export declare function isDiscoveryEnabled(): boolean;
 export declare function getMaxInterviewQuestions(): number;
+export declare function isPonytailModeEnabled(): boolean;

package/dist/src/pil/config.js

@@ -11,6 +11,21 @@ export function isUnifiedPilEnabled() {
         return true;
     return false;
 }
+/**
+ * MUONROI_PIL_UNIFIED_BUDGET_MS: client-side timeout for the unified
+ * /api/pil-context call in Layer 1 Pass 3. The server's uncached path is
+ * classifier-bound (~1.5-3s; its own classifier timeout is 3500ms), so the
+ * legacy 1500ms budget aborted before the server could ever answer — the
+ * unified path then ALWAYS fell back to legacy, making MUONROI_PIL_UNIFIED=1 a
+ * no-op. Default 3500ms clears realistic uncached latency; cache hits return in
+ * ~1ms regardless. Clamped to [1000, 8000].
+ */
+export function getUnifiedPilBudgetMs() {
+    const raw = Number(process.env.MUONROI_PIL_UNIFIED_BUDGET_MS);
+    if (!Number.isFinite(raw))
+        return 3500;
+    return Math.max(1000, Math.min(8000, Math.trunc(raw)));
+}
 /**
  * MUONROI_LLM_FIRST_CLASSIFY: model-first Layer-1 classification. When enabled
  * (default), the configured model classifies taskType/intentKind/style at the
@@ -22,6 +37,17 @@ export function isUnifiedPilEnabled() {
 export function isLlmFirstClassifyEnabled() {
     return process.env.MUONROI_LLM_FIRST_CLASSIFY !== "0";
 }
+/**
+ * G3 (b1): on the model-first classify path, also fetch the unified pil-context
+ * brain so layer3 renders the rich source="unified" injection (server-attributed
+ * t0/t2 + t1_rules + the rateable ledger) instead of falling through to its
+ * legacy dense-only /api/search round-trip. Default ON; set "0" to revert to
+ * the legacy layer3 injection on the default path. Independent of
+ * MUONROI_PIL_UNIFIED (which gates the OFFLINE cascade's unified pass).
+ */
+export function isLlmFirstBrainEnabled() {
+    return process.env.MUONROI_LLM_FIRST_BRAIN !== "0";
+}
 export function isDiscoveryEnabled() {
     return process.env.MUONROI_PIL_DISCOVERY !== "0";
 }
@@ -33,4 +59,13 @@ export function getMaxInterviewQuestions() {
     const v = Number(process.env.MUONROI_PIL_MAX_QUESTIONS);
     return Number.isFinite(v) && v >= 1 && v <= 5 ? v : 3;
 }
+export function isPonytailModeEnabled() {
+    if (process.env.MUONROI_PONYTAIL_DISABLE === "1")
+        return false;
+    if (process.env.MUONROI_PONYTAIL_DISABLE === "0")
+        return true;
+    if (process.env.VITEST === "true" || process.env.NODE_ENV === "test")
+        return false;
+    return true;
+}
 //# sourceMappingURL=config.js.map